sgugger
Follow
sgugger
Follow
Popular repositories
3,004 contributions in the last year
Activity overview
Contributed to
huggingface/transformers,
huggingface/doc-builder,
huggingface/accelerate
and 13 other
repositories
Contribution activity
May 2022
Created 28 commits in 3 repositories
Created a pull request in huggingface/accelerate that received 52 comments
Big model inference
This PR introduces a set of tools to enable inference on large models that do not fit fully on one GPU, or even in RAM. Namely a context manager t…
+2,796
−2
•
52
comments
Opened 13 other pull requests in 1 repository
huggingface/transformers
1
open
12
merged
- Automatically sort auto mappings
- Fix Trainer for Datasets that don't have dict items
- Handle copyright in add-new-model-like
- Black preview
- Fix template init
- Remove torchhub test
- Make sure telemetry arguments are not returned as unused kwargs
- Deprecate model templates
- Remove Python and use v2 action
- Fix RNG reload in resume training from epoch checkpoint
- Make Trainer compatible with sharded checkpoints
- Clean up setup.py
- Disable Flax GPU tests on push
Reviewed 90 pull requests in 4 repositories
huggingface/transformers
25 pull requests
- Updated checkpoint support for Sagemaker Model Parallel
- Guide to create custom models in Spanish
- Automatically sort auto mappings
- install dev. version of accelerate in docker file
- Add trajectory transformer
- Adds support for OPT in Flax and TF.
- BLOOM
- Traced models serialization and torchscripting fix
- Added translation of installation.mdx to Portuguese Issue #16824
- OPT - fix docstring and improve tests slighly
- Add XLNet OnnxConfig
- Add Tensorflow Swin model
- Add test to ensure models can take int64 inputs
- Update data2vec.mdx to include a Colab Notebook link (that shows fine-tuning)
- Remove duplicated os.path.join in Trainer._load_rng_state
- ViT and Swin symbolic tracing with torch.fx
- Remove columns before passing to data collator
- Add OPT
- Ensure tensors are at least 1d for pad and concat
-
Remove unnecessary columns for all dataset types in
Trainer - Fix LED documentation
- [WIP] Enable reproducibility for distributed trainings
- Mlflowcallback fix nonetype error
- Add DebertaV2ForMultipleChoice
- Add MLFLOW_FLATTEN_PARAMS support in MLflowCallback
- Some pull request reviews not shown.
huggingface/accelerate
14 pull requests
- Remove tensor call in save_state for XLA
-
Fix minor typo in
notebook_launchererror message - fix tracking
- Handle deprication errors in launch
- Big model inference
- DeepSpeed and FSDP plugin support through script
- Fixing metric eval in distributed setup
- Fix sample calculation in all examples
- Fix prompt for num_processes
- Better prompt for number of training devices
- Handle Manual Wrapping in FSDP. Minor fix of fsdp example.
- Improve num_processes question in CLI
- Refactor utils into its own module
- Fixed a typo to enable running accelerate correctly